Self-supervised learning for characterising histomorphological diversity and spatial RNA expression prediction across 23 human tissue types

Abstract

As vast histological archives are digitised, there is a pressing need to be able to associate specific tissue substructures and incident pathology to disease outcomes without arduous annotation. Such automation provides an opportunity to learn fundamental biology about how tissue structure and function varies in a population. Recently, self-supervised learning has proven competitive to supervised machine learning approaches in classification, segmentation and representation learning. Here, we leverage self-supervised learning to generate histology feature representations using 1.7M images across 23 healthy tissues in 838 donors from GTEx. Using these representations, we demonstrate we can automatically segment tissues into their constituent tissue substructures and pathology proportions, and surpass the performance of conventionally used pre-trained models. We observe striking population variability in canonical tissue substructures, highlight examples of missing pathological diagnoses, incorrect assignment of target tissue and cross-tissue contamination. We demonstrate that this variability in tissue composition leads to a likely overestimation of eQTL tissue sharing and drives dramatic differential gene expression changes. We use derived tissue substructures to detect 284 tissue substructures and pathology specific eQTLs. As our derived histology representations are rich morphological descriptors of the underlying tissue, we introduce a multiple instance learning model that can predict and spatially localise individual RNA expression levels directly from histology to specific substructures and pathological features. We validate our RNA spatial predictions with matched ground truth immunohistochemistry (IHC) for several well characterised marker genes, recapitulating their known spatial specificity. Finally, we derive a gene expression spatial enrichment metric, allowing us to detect genes specifically expressed within sites of pathology (e.g. arterial calcification). Together, these results demonstrate the power of self-supervised machine learning when applied to vast histological datasets to allow researchers to pose and answer questions about tissue pathology, its spatial organisation and the interplay between morphological tissue variability and gene expression.